How to estimate component wear

Abstract

The present disclosure relates to a computer-implemented method for estimating component wear, where at least one clustering analysis can be applied to measurements taken during use of the component to identify clusters and provide alerts.

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method for estimating wear of a component, wherein at least one clustering analysis is applied to measurements taken during use of the component to identify clusters and issue alerts. [Background technology]

[0002] As machine components are used, they undergo partial wear. Wear involves deterioration or a change in shape of the component and often occurs on surfaces in contact with another part of the machine or with an external surface. Wear can be chemical, mechanical, or both, and can be exacerbated by conditions surrounding the contact, such as pressure, temperature, and lubrication.

[0003] Wear requires that components be inspected and replaced, as continued use of worn components can be inefficient and unsafe.

[0004] In a vehicle, for example, parts that are commonly replaced include windshield wiper blades, filters, brake pads, tires, belts, and cylinders. All of these components are susceptible to wear due to friction and heat. Wear in any one of these components can make the vehicle less safe to operate, both because the component is less effective as it wears (e.g., worn brake pads result in poor stopping capabilities), and because eventual failure of the component can result in the vehicle losing important functionality (e.g., a broken windshield wiper blade can cause the driver to lose visibility in adverse weather conditions).

[0005] Wear is also detrimental during machining processes, such as metal cutting, in which the relative motion of a tool and a workpiece removes material in the form of chips to create a net shape. Indeed, one of the most significant yet unavoidable machining challenges is the continuous wear on the tool. In this case, tool wear is the removal of material from the cutting surface due to the interaction between the tool and the workpiece. Tool wear is known to cause many problems in machined components, including shape and dimension inconsistencies, vibration and chatter, and poor surface finish.

[0006] Machining with worn or damaged tools can result in surface defects on the component and crack propagation over time, resulting in the component being unusable or being discarded prematurely. As a result, when machining safety-critical components, the industry employs time-based maintenance methods and discards tools at set intervals, regardless of their wear state. Typically, tools may be discarded with 50-80% of their useful life remaining.

[0007] Evaluating component wear during use can be challenging because the component may be inaccessible and / or moving, preventing visual inspection from providing useful results. Using a vehicle as an example, inspecting brake pad wear may require removal of the wheels, a task that the average vehicle user may not be able to perform. Belts and other internal combustion engine components can be even more difficult to inspect.

[0008] Accessibility also poses challenges during machining. Attempting to inspect components used for machining, such as tools, has the added drawback of requiring the tool to be taken out of service while the inspection is performed. During the design and testing phase of a tool, designers may opt for offline wear measurements in a "run-to-failure" approach. This requires intermittent interruptions to the cutting process and temporary removal of the tool from the machine to complete wear measurements using an optical microscope. This is a time-consuming and expensive method, and it also introduces confounding effects on the process, such as differences in the tool's position during a first machining cycle compared to a second machining cycle. Because the difference in position affects cutting conditions, wear evolution may differ between machining cycles of the same duration.

[0009] 1 shows a typical breakdown of time spent in various states during the life of a tool. Less than half of the life of this exemplary typical tool is spent machining, with the remainder of the tool life wasted due to performing either workpiece or tool measurements.

[0010] TCM aims to shift the industry's focus from a preventative approach to predictive maintenance strategies in order to reduce waste. To increase process efficiency, a TCM system should ideally help reduce time spent measuring and increase machining time.

[0011] Supervised machine learning can be used to assist TCM techniques. As with any supervised learning technique, data associated with a large number of previously used tools can be used to build a model that attempts to define the relationship between the measured degradation of previously used tools and the tool's characteristics. For example, a simple model created by supervised learning techniques can find and define the relationship between machining time to failure and tool diameter, workpiece hardness, and tool revolutions per minute (rpm). Such a TCM technique can predict time to failure, in this case for a new tool, using tool diameter, workpiece hardness, and tool revolutions per minute (rpm) as input parameters to the machine learning model. This still means that TCM can be inaccurate due to the complex nature of machining, various operating conditions, and the availability of descriptive labels for wear states. Summary of the Invention [Problem to be solved by the invention]

[0012] Improved assessment of component wear is desired to reduce component downtime and increase safety. [Means for solving the problem]

[0013] The invention is defined by the accompanying independent claims. Embodiments of the invention are set out in the dependent claims.

[0014] In a first aspect, there is provided a computer-implemented method for estimating wear of a component, comprising: performing a first clustering analysis on a first plurality of measurements to identify one or more clusters in the first measurements, the first measurements comprising measurements of a parameter of the component taken while the component is in use; and issuing an alert if a new cluster is identified in the first measurements.

[0015] In this way, the method provides an estimate of component wear by highlighting points in time when there has been enough change in the measured parameters to warrant a new cluster being identified.

[0016] The present method uses measurements taken during the use of the relevant component, thereby avoiding the deleterious effects associated with the use of a test component or the use of data related to a component other than in use. The present method is therefore more accurate because it does not rely on the assumption that components wear out during use in a manner similar to that used to build or train wear models, e.g., trained classifiers. The wear evolution of a component can vary widely depending on the component category and its condition at the time of use, but also on the random nature of the component, properties that are not easily measured or accurately predicted due to, for example, contamination and microstructure. Therefore, utilizing the assumption of uniform tool characteristics can reduce the accuracy of wear estimates. In contrast, the present method uses current data measured from the specific tool whose wear needs to be estimated. A more accurate estimate of component wear can reduce component downtime by avoiding unnecessary inspections, increase safety by indicating the need for inspection when other estimates fail to do so, and reduce the remaining wasted life of a tool at the time of tool disposal.

[0017] Clustering analysis, which is performed while the tool is in use and updated as new measurements are taken, is advantageous because components often exhibit significant changes in their behavior when they reach a certain wear level. For example, a crack may appear on a machining tool or a fray may begin in a drive belt, causing a change in acoustic emissions from the component, a change in forces around the component, or a change in temperature at the friction zone where the component contacts the tool. Clustering analysis can identify such changes in measurements. Furthermore, clustering analysis is advantageous over implementing simple threshold measurements that can trigger alerts because threshold-based alerts are susceptible to being triggered by outliers. For example, a very brief spike (e.g., a single measurement) in the temperature between the drive belt and sprocket may be caused by factors unrelated to belt wear (e.g., debris quickly entering or exiting the friction zone). Clustering analysis would not consider this outlier to be the start of a new cluster and thus would not trigger an alert that would result in unnecessary inspection, maintenance, or component scrapping.

[0018] Issuing a warning may be an alert to a component and / or associated machine controller to terminate operation in preparation for inspection or as a safety measure. Issuing a warning may further include generating an audible or visual signal to an operator or a warning sent to an external device as a notification.

[0019] The method may include performing a second clustering analysis on a second plurality of measurements to identify one or more clusters in the second measurements, the second measurements including measurements of parameters of the component taken during further use of the component after the alert was generated, and the method may further include issuing an alert if new clusters are identified in the second measurements.

[0020] After initial inspection of the component with a warning, the component may be deemed suitable for further use and returned to its associated machine. During further use, a second clustering analysis is performed in a similar manner as the first clustering analysis, and a warning is issued in a similar manner. The second clustering analysis benefits from the same advantages as the first clustering analysis.

[0021] The method may further include continuing the first clustering analysis on the first measurements and introducing the second measurements into the first clustering analysis to identify one or more clusters in the combined first and second measurements, and issuing an alert if a new cluster is identified in the combined first and second measurements.

[0022] Clustering analysis offers advantages both when performed over the entire life of a component (global analysis) and when only measurements taken since the most recent component removal from the associated machine, for example during inspection, are used (local analysis). The technological advancement in global analysis can provide accurate estimates of overall component wear, i.e., the component's remaining life. This is due to clustering analysis's ability to access measurements of components under a variety of conditions, including when new and over time. However, global analysis is less susceptible to changes in component behavior that are significant on a local scale, perhaps important enough to warrant inspection, but not significant in relation to the component's overall lifespan. Global analysis is particularly immune to significant changes in wear characteristics that manifest themselves in measurements, especially when the measurements taken after inspection are significantly different from those taken before inspection due to differences in position after inspection.

[0023] By performing global and local analyses in parallel, alerts can be generated for emerging cluster indications, thus enabling more accurate estimation of component wear. Alerts generated by the first and second clustering analyses can be distinguished from one another because they can be identified by either the controller or the operator. For example, if local analysis is used to monitor when an inspection is needed, the alert may be a notification to the operator to recommend an inspection. An alert generated by the global analysis may be a notification to both the operator and the manager that a tool is nearing the end of its life.

[0024] The first clustering analysis may be referred to herein as a global clustering analysis, and the second clustering analysis may be referred to herein as a local clustering analysis. It will be appreciated that although only two parallel analyses are described (one including lifetime measurements and the other including measurements since the last inspection), any number of parallel analyses may be performed in accordance with embodiments of the present invention. For example, a clustering analysis may be performed based on measurements between the second and fifth inspection of a component.

[0025] The first and / or second clustering analysis may be performed by an unsupervised machine learning algorithm, and the first and / or second clustering analysis may not prescribe the number of clusters to be determined during the analysis.

[0026] Thus, tool wear estimation is again accurate because the method does not utilize a manually selected number of clusters, which does not reflect the behavior of the component during use. The predetermined number of clusters instructs the clustering analysis on how many clusters it should place data points (in this case, measurements). This number can be determined based on the number of clusters that have led to effective and accurate estimations for similar components used previously. However, this has the same problems discussed above regarding differences, sometimes imperceptible, between components that can limit the accuracy of predictions based on prior knowledge.

[0027] The unsupervised machine learning algorithm is the Dirichlet process mixture model (DPMM). The properties of the DPMM have been found to be particularly suitable for estimating component wear, especially when global and local DPMMs are implemented in parallel.

[0028] The warning should be a recommendation to inspect the component.

[0029] The parameter of the component may be one or more of the temperature of the component, the force applied to the component, and the acoustic emissions from the component. In this way, i.e., by taking measurements of a parameter other than a direct measurement of the wear itself, measurements can be performed while the component is in use. Thus, removal of the component from its associated machine can be avoided, thereby reducing dead time for the component and eliminating the possibility of misplacement.

[0030] The parameter may be an acoustic emission from the component, and the first and / or second clustering analysis may use a plurality of adjacent frequency ranges as candidate features.

[0031] Acoustic emission has been found to be a particularly effective parameter to measure when analyzing measurement data using clustering analysis. Unsupervised clustering analysis advantageously does not utilize a model trained on training data and optionally does not have a predetermined number of clusters, although it may be advantageous to provide some guidance related to candidate features for the algorithm. Candidate features are variables of interest to the algorithm; therefore, in the case of unsupervised clustering algorithms, candidate features are variables with which the data points to be clustered are associated. Frequency bins are a useful candidate feature when the variable in question is the acoustic emission of a component.

[0032] The component may be a machining tool. As will be appreciated, in a preferred embodiment, the component is a machining tool, and the methods for estimating component wear described herein are suitable and advantageous for any component of a machine, particularly any component of a machine that, in use, produces a parameter suitable for indirect measurement, including, but not limited to, temperature, sound, force, pressure, any wavelength of light, etc. An example related to a vehicle component is also given, but again, this is not intended to limit the invention.

[0033] Thus, the associated machinery referred to herein may include the workpiece, robotic control components, motors, etc., i.e., components other than the tool itself used to machine the workpiece. In the previous vehicle example, if the component is a drive belt, the associated machinery may include the driven component, e.g., an alternator or pump, a sprocket to which the belt is attached, etc.

[0034] Each frequency range may correspond to a predetermined harmonic of chip formation in the tool, and optionally each frequency range may be spaced 4 kHz apart. Chip formation is a particularly relevant factor when the component in question is a machining tool, as the long sliding distances and high rpm of the machining tool make it particularly susceptible to chip or crack formation. By using frequency ranges based on chip harmonics, new clusters are more likely to be identified when a new harmonic of chip formation is reached, i.e., when a chip becomes increasingly likely to occur.

[0035] In a second aspect, there is provided a method of inspecting a component for wear, the method comprising: performing a computer-implemented method according to the first aspect; and in response to a warning output from the computer-implemented method, inspecting the component for wear.

[0036] In a third aspect, there is provided a data processing apparatus, characterized in that it comprises means for performing the steps of the first aspect. The data processing apparatus may be any processing means capable of or configured to perform the steps of the methods described herein. The data processing apparatus may be a single processing unit, a wired distributed system, or a wireless distributed system. Alternatively, the data processing apparatus may host a cloud server on which the analysis is performed.

[0037] In a fourth aspect, there is provided a computer program having instructions embodied therein that, when executed by a computer, cause the computer to perform the steps of the first aspect. The computer program may, for example, be a mobile application configured to run on a mobile device.

[0038] In a fifth aspect, there is provided a computer-readable storage medium having instructions embodied thereon that, when executed by a computer, cause the computer to perform the steps of the first aspect.

[0039] The present invention can be more fully understood when considered in connection with the accompanying drawings, in which like reference numerals designate the same or similar components throughout the several views. [Brief explanation of the drawings]

[0040] [Figure 1] FIG. 1 shows a breakdown of time spent in a typical tool wear test setup. [Figure 2] FIG. 1 illustrates the cast sample collection process. [Figure 3] FIG. 1 illustrates a machining setup according to one embodiment of the present invention. [Figure 4] FIG. 1 shows spectrograms of a PcBN tool under specific cutting conditions. [Figure 5] FIG. 10 shows DPMM clustering results for test tool A1 using harmonics of chip generation frequency from AE as input features. [Figure 6] FIG. 10 shows DPMM clustering results for test tool A2 using harmonics of chip generation frequency from AE as input features. [Figure 7] FIG. 10 shows DPMM clustering results for test tool A3 using harmonics of chip generation frequency from AE as input features. [Figure 8] FIG. 10 shows DPMM clustering results for test tool A4 using harmonics of chip generation frequency from AE as input features. [Figure 9] FIG. 10 shows DPMM clustering results for test tool B1 using harmonics of chip generation frequency from AE as input features. [Figure 10]FIG. 10 shows DPMM clustering results for test tool B2 using harmonics of chip generation frequency from AE as input features. [Figure 11] FIG. 10 shows DPMM clustering results for test tool B3 using harmonics of chip generation frequency from AE as input features. [Figure 12] FIG. 10 shows DPMM clustering results for test tool B4 using harmonics of chip generation frequency from AE as input features. [Figure 13] FIG. 10 shows DPMM clustering results for test tool A5 using harmonics of chip generation frequency from AE as input features. [Figure 14] FIG. 10 shows DPMM clustering results for test tool A5 using harmonics of chip generation frequency from AE as input features. [Figure 15] FIG. 10 illustrates the number of clusters formed for a range of α values ​​for two example tools (right-A4 and left-B3). [Figure 16] 1 illustrates a system 100 according to one embodiment of the present invention. [Figure 17] FIG. 10 shows DPMM clustering results for test tool A1 using harmonics of chip generation frequency from AE as input features. [Figure 18] FIG. 10 shows DPMM clustering results for test tool A2 using harmonics of chip generation frequency from AE as input features. [Figure 19] FIG. 10 shows DPMM clustering results for test tool A3 using harmonics of chip generation frequency from AE as input features. [Figure 20] FIG. 10 shows DPMM clustering results for test tool A4 using harmonics of chip generation frequency from AE as input features. [Figure 21] FIG. 10 shows DPMM clustering results for test tool B1 using harmonics of chip generation frequency from AE as input features. [Figure 22]FIG. 10 shows DPMM clustering results for test tool B2 using harmonics of chip generation frequency from AE as input features. [Figure 23] FIG. 10 shows DPMM clustering results for test tool B3 using harmonics of chip generation frequency from AE as input features. [Figure 24] FIG. 10 shows DPMM clustering results for test tool B4 using harmonics of chip generation frequency from AE as input features. [Figure 25] FIG. 10 shows DPMM clustering results for test tool A5 using harmonics of chip generation frequency from AE as input features. [Figure 26] FIG. 10 shows DPMM clustering results for test tool A5 using harmonics of chip generation frequency from AE as input features. DETAILED DESCRIPTION OF THE INVENTION

[0041] A detailed example in the field of machining is provided to demonstrate and explain embodiments of the present invention. In this example, the component is a machining tool and the measured parameter is acoustic emission. It should be understood that this is merely an example, and methods and systems according to embodiments of the present invention can be applied to any number of technical fields where components wear down through use, exposure, or any other means. Similarly, parameters other than acoustic emission that are measured indirectly during use of the component, preferably without interruption of use, can be used as input data for a clustering algorithm that describes the components.

[0042] For example, a clustering analysis can be performed on temperature measurements taken near a fan (a component) in a desktop computer (related machine). A clustering algorithm, as described in detail herein, can be applied to the temperature measurements, and an alert can be issued when new clusters are identified. The alert can, for example, prompt replacement or inspection of the fan.

[0043] We now describe a realistic TCM system using Dirichlet Process Mixture Models (DPMMs) that can handle confounding effects while maintaining inference accuracy for previously unseen tools (typically either novel tools never used for machining or tools never directly measured to obtain measurements of tool wear). The resulting system is therefore potentially more versatile than previous attempts and easier to adopt in industry.

[0044] An unsupervised clustering approach using a Dirichlet process mixture model is used to detect and diagnose changes in cutting process characteristics online. In addition to providing a useful monitoring tool, this approach potentially reduces the need for relevant, comprehensive wear measurements for prediction. The model is well-suited to the irregular and unpredictable nature of tool wear evolution, since the number of clusters required to determine possible damage states is not set a priori. As a result, the method is adapted to handle variations across homogeneous and heterogeneous groups of tool material compositions.

[0045] The proposed approach is presented here as a way to reduce the time required for trials to characterize the wear characteristics of new tools. In the illustrated example, results show that this approach results in an almost 30% reduction (on average) in test time during external turning of case-hardened steel.

[0046] The ability to monitor and predict tool degradation during machining is an important objective, as the wear stage has a significant impact on the surface quality of the machined component. However, to build a comprehensive condition monitoring system for diagnosis and prognosis, extensive measurements and knowledge of tool wear are required. Collecting labeled datasets containing damage information for this purpose can be expensive and time-consuming.

[0047] In this study, DPMM is used as an unsupervised method to detect characteristic changes in data for online damage detection. The inventors apply DPMM to acoustic emission (AE) datasets collected during grinding operations. DPMM enables clustering as data is being collected, without the need to set an a priori number of possible clusters, reducing the need for extensive prior knowledge of the machining process. The need for pre-labeled training data is also eliminated, reducing the costs associated with data collection.

[0048] This study solves the problem associated with confounding effects introduced by varying the work during a trial by using two DPMMs in parallel. The invention, in this case, uses a new cluster start to prompt an intervention, namely manual inspection of the tool. This was expected to result in a reduction in the number of process interruptions. The objective was to increase the time spent on machining and decrease the time spent on tool measurement. By tracking the characteristics of the data through time, another goal was to stop the machining process before breakage and reduce the number of early tool disposals. Although not included in this study, it may be possible to employ a semi-supervised approach by adding damage labels to the model, thereby reducing the number of false positives.

[0049] Dirichlet process

[0050] DPMMs are used in many research fields to cluster data. In the field of natural language processing, translation and text generation research has seen researchers use DPMMs to cluster verbs, using this technique to organize and predict words. In medical research, DPMMs have been used to classify brain tissue from magnetic resonance imaging (MRI) scans. While GMMs work effectively for well-defined tissue patterns, DPMMs do not require a priori setting of the number of clusters, allowing researchers to classify abnormal brain data.

[0051] An introduction to Dirichlet Process Gaussian Mixture Models is given in this section, where Gaussian distributions are used as the underlying distributions.

[0052] DPMM can be used to cluster Gaussian and non-Gaussian distributed data. DPMM can be viewed as an infinite Gaussian mixture model (IGMM), in which a mixture of Gaussian distributions is used to cluster the data. Here, the number of Gaussian distributions employed can tend to infinity, making it possible to model any non-Gaussian distributed data set. That is, a cloud of data that does not follow a Gaussian distribution can be divided into an infinite number of small clusters, which are themselves Gaussian distributed. IGMM can also be used to learn information about the probability of each data point belonging to each cluster. The idea is to find clusters in the data while also finding the parameters (size, shape) and labels of these clusters (cluster 1, ..., cluster n). Finding the labels and cluster parameters in one step is not possible, so Gibbs sampling is used in this study to estimate the joint probability distribution. It is not a requirement that the DPMM have knowledge of the number of clusters that may exist in the data before applying the algorithm; i.e., the Gibbs sampler can start a new cluster if the data already evaluated by the algorithm is sufficiently different from the current data point.

[0053] The generative model of DPMM is shown in Equations (1a) to (1e). In the Gaussian mixture model, each data point xi (i = 1, ... N, where N is the number of data points) is sampled from a Gaussian distribution (Equation (1a)), and clusters of data are labeled c i Assume that we have

[0054] For each Gaussian distribution, we obtain a closed-form solution for the posterior distribution using a conjugate prior (normal inverse Wishart (N|W) distribution on the mean (Equation (1b)) and covariance (Equation (1c)) with hyperparameters (μ, Σ, κ, ν). The data is normalized, and the prior clusters have zero-mean and unit-variance Gaussian distributions in this study. JPEG0007749686000001.jpg62116

[0055] The cluster labels are sampled from a multinomial distribution (Equation (1d)). The mixture probability π is the probability of data belonging to each cluster. To calculate these probabilities, the Dirichlet distribution is used because it is the conjugate prior of the multinomial distribution (Equation (1e)). π is controlled by the strength parameter α of the Dirichlet process prior (α is sometimes called the scaling or scale parameter). The scaling parameter α determines the probability of identifying a new cluster; that is, the probability that a new cluster is identified for a new data point rather than being assigned to an existing cluster is proportional to α.

[0056] Ultimately, the goal of this process is to find a posterior distribution over the cluster labels from which to select the most likely labels. Finding the probability of all cluster labels given all the data (equation (2)) is extremely difficult, since the cluster parameters (μ ci , Σ ci ), while simultaneously sampling each cluster to find the mixture ratios of all clusters. p(c | X) (2)

[0057] A collapsing Gibbs sampler can be implemented to find this solution. The collapsing Gibbs sampler sequentially samples a new set of cluster parameters based on the sample of labels and a new set of labels based on the parameters (Equation (3)). Figure 2 shows the steps followed by the Gibbs sampler. Each of these sampling steps is preferably implemented in closed form for the selection of conjugate prior distributions. In other words, it is an efficient Markov Chain Monte Carlo method to sequentially find distributions for cluster labels and distributions for cluster parameters. The collapsing Gibbs sampler is implemented on a window of data, the length of which is specified according to the available computational power. p(c i | x i , X -i , c -i ) (3)

[0058] Equation (3) is the first probability distribution of interest. It is the distribution that the model uses for all data X -i , all other clusters c -i , and assuming we have a new data point, the cluster label c i This posterior probability has a multinomial distribution. In this case, -i is the data point i This refers to all data points except for

[0059] To compute this, the Gibbs sampler randomly assigns data to clusters, then sequentially removes data points, updating the cluster parameters and finding the cluster that best fits the data point. The prior likelihood of drawing data from existing clusters (k=1,...,K) can be found in Equation (4), where N -i,k is the number of data points in the current class, and N is the total number of data points once the one being considered has been removed. JPEG0007749686000002.jpg14127

[0060] The new cluster (k * ) is calculated using equation (5). JPEG0007749686000003.jpg14129

[0061] The posterior predicted likelihood of a point belonging to each cluster should now be calculated, which is the probability of assigning the current data point to cluster k, given the value of the data, the current cluster it is in, all of the data already in that cluster, and some hyperparameters.

[0062] To calculate the likelihood of a DPMM, equation (6) is used, which is called the posterior predictive distribution. This means that the likelihood that a data point occurs in a cluster defined by the posterior distribution on the parameters, in this case D k is all the data the algorithm has seen in a given cluster. For the DPMM model, this is represented by a multivariate t-distribution, which has heavier tails than a Gaussian distribution, making small clusters more receptive to new data points. TIFF0007749686000004.tif12163

[0063] The posterior distribution is the marginal likelihood, i.e., the probability that this is a valid probability distribution, and the cluster label c of point i i The posterior distribution is normalized by the sum of all the computed priors to find the multinomial distribution for . If a point is assigned to a new cluster, the NIW prior is used to initialize the cluster and the number of clusters is increased by 1. This process is repeated until all data within the window has been evaluated again.

[0064] There are three major advantages to using this model. First, the model does not require the operator to set the number of clusters (or in this case, the damage state) before using the algorithm, eliminating the need for prior process knowledge, which can be difficult to obtain. Second, the entire model is controlled by hyperparameters, so threshold tuning and calibration are not required. Third, DPMM allows the covariance function to vary with the input data, resulting in a model that can handle dissimilar data sets, which can be useful when detecting damage to multiple tools that wear differently from each other.

[0065] Experimental setup

[0066] The test setup used here to explore the application of DPMM is shown in Figure 3, where a PcBN tool is used to machine a case-hardened workpiece until catastrophic failure. These are referred to as "end-of-life" tests. Because it is not possible to measure tool wear during machining, these tests are segmented to perform four consecutive cuts, called passes, of the workpiece at a time, followed by tool inspection. The inspection consists of removing the tool from the machine and measuring the flank and crater wear using a 2D optical microscope and a 3D scanning microscope, respectively. On average, it takes more than eight minutes to complete the four passes and associated tool wear measurements. Over 42% of the time spent on tool wear testing is spent on these measurements and related activities.

[0067] The setup has limitations that must be considered when selecting a model for TCM. First, removing the tool for measurement is unavoidable because it is the only way to collect damage labels. Once the tool is removed, it cannot be returned to the exact position before the measurement and fixed. As a result, the depth of cut value is uncertain for the next pass. Depth of cut is crucial to the mechanics of the machining, and changing the depth of cut can affect the forces, stresses, and dynamics of the process. Second, during the four-pass interval, the workpiece is replaced with a new one. This is because the case hardening penetrates only a thin layer around the periphery of the workpiece, and it is likely to be removed by machining after four passes.

[0068] In this case, acoustic emissions were measured using a Mistras Micro-30D differential sensor with a sampling rate of 1 MHz. Taking the Nyquist criterion into account, this results in a frequency range of 0-500 kHz. An Okuma Space Turn lathe LB3000 EXII was used for this experiment. Cuttings were collected in an attempt to further explore the mechanism of AE generation.

[0069] AE signals were collected from different tool types, with the only variation being the material composition, or "grade" of the material. Tool grade specifies the percentage of cubic boron nitride (cBN) particles within that composition. The blade has a direct effect on the tool's behavior and defines its suitability for the type of machining operation. The two grades used in this study are referred to as Grade A (Tools A1-A5) and Blade B (Tools B1-B5).

[0070] Feature Selection for DPMM

[0071] AE signals can be visualized in the frequency domain as a spectrogram. In Figure 4, a spectrogram for an example tool is displayed as a function of sliding distance, which is the distance traveled by the tool. In Figure 4, frequency is plotted in kilohertz on the y-axis, sliding distance is plotted in kilometers on the x-axis, and shading represents the power of the signal as average power per unit bandwidth in decibels per hertz. In this case, the critical frequency ranges are highlighted, where sawtooth generation is hypothesized to occur at the tip (also referred to in this study as the chip generation frequency) and possibly at harmonics of these frequencies due to nonlinearities in the system.

[0072] Finally, DPMM has the potential to automatically cluster data in an industrial setting when operating online. To achieve this, the input features should behave similarly across a homogeneous set of tools. It is clear from the spectrogram that the power of chip-forming frequencies increases in intensity with the progression of tool wear. As a result, four frequency bins (spaced 4 kHz apart) that carry harmonics of chip-forming frequencies (throughout tool life) are selected as candidate features for each tool grade. Prior knowledge of the process is required to identify these features that increase in intensity at the end of tool life. This behavior cannot be identified a priori for new tools when the process is operating online. Here, the frequency bands used reduce the requirement for specific prior knowledge, but the identification of these bands is achieved using preliminary trials on any new tool grade.

[0073] As will be appreciated, similar preliminary trials can be performed on any component of a machine to determine candidate features. Such preliminary trials need not focus on acoustic emissions, but rather on any measurable parameter described herein. Furthermore, the clustering algorithms described herein can be performed without using candidate features determined from preliminary trials. For example, arbitrary ranges can be used during initial trials of the component and then iteratively refined. By way of example only, if temperature is selected as the parameter to be measured during use of the component, a temperature range beginning at 50°C and increasing in increments of 25°C, 50°C, 75°C, or 100°C can be used. Similarly, force, emission wavelength, and pressure, with or without preliminary testing, can use ranges in Newtons, wavelengths, and Pascals as candidate features.

[0074] Parallel DPMM

[0075] Clusters in DPMM form due to changes in the mean and variance of the input features.

[0076] During machining, changes in the mean and variance of AE generation are assumed to be caused by the onset of tool wear. However, unavoidable confounding influences, such as changes in the workpiece at specified intervals or changes in the positioning of the tool post measurement, can also cause variations in the AE signal. It is important to distinguish between these two causes of cluster generation when using DPMM to avoid false positives.

[0077] In structural health monitoring (SHM), confounding effects can manifest themselves as short-term and long-term trends, which must be distinguished from damage for accurate prediction. Notably, principal component analysis can also be used as a technique in dealing with confounding effects, where principal components that exhibit low variance can be used as damage-sensitive features and are not compromised by high-profile environmental issues. While these techniques are successful in suppressing unwanted trends, they require a priori representative sets of training data.

[0078] In this study, it is proposed to take into account the confounding effects and also to preserve the susceptibility to possible tool damage. Here, it is proposed to operate two DPMMs in parallel. 1. A first DPMM is started at the beginning of each tool test and runs throughout the entire tool life. This DPMM is hereafter referred to as the global DPMM. 2. A second DPMM is started at the beginning of each tool test and reset at the beginning of each of the four passes of the workpiece, i.e., wherever the workpiece changes and tool positioning changes occur. This DPMM is called the local DPMM.

[0079] The next section considers the application of parallel DPMM to the above dataset.

[0080] Experimental results

[0081] For each tool, the above features are input into two parallel DPMMs, with the Gibbs sampler window length set to 200 (200 seconds of data) to allow convergence to the target distribution.

[0082] Figures 5-12 show the results for tools A1-A4 and B1-B4. Figures 5-8 show the parallel clustering for tools A1-A4, respectively, and Figures 9-12 show the parallel clustering for tools B1-B4, respectively. Each of these figures includes results from a global DPMM and a local DPMM. In all cases, the top four plots relate to the global DPMM, and the bottom four plots relate to the local DPMM. Harmonics of the chip generation frequency from acoustic emissions are used as input features F1-F4 in each case. As mentioned above, features F1-F4 are frequency "bins"; in this example study, these bins are spaced 4 kHz apart and carry the chip generation frequency harmonics throughout the tool life. F1 is the lowest Hz frequency bin, and F4 is the highest Hz frequency bin. The y-axis scale for each feature was normalized before plotting.

[0083] In the methods shown in Figures 5-14 and 17-26, new clusters are identified in both the global and local DPMMs. New clusters in the global DPMM are identified by vertical lines, while in the local DPMMs they are identified by a change in color. In both cases, once a new cluster is identified for any one of the candidate / input features F1-F4, new clusters are initiated for all other candidate features. For example, in Figures 10 and 22, points of interest 10.1 and 22.1 indicate that a new global cluster is identified based on the data point in the fourth frequency bin (feature F4). Features F1-F3 are updated based on this identification, and these features initiate classification of the data point into a new cluster (see points of interest 10.2 and 22.2), even though these data points likely do not meet the new cluster identification threshold.

[0084] The local clusters are homogenized across all input features in the same manner. However, unlike global clusters, the local clustering analysis starts anew with every tool inspection. In other words, the local clustering graph includes multiple local clustering analyses, each using a data set that includes the data points between the vertical dashed lines.

[0085] It will be appreciated that this method of applying cluster identification based on one feature to all other features in the associated DPMM (i.e., in the global DPMM or in the local DPMM) is one possible embodiment, but other methods are possible in accordance with the present invention. The features could be kept completely separate, and a new alert could be issued for each feature individually. Alternatively, a threshold number of features could be implemented to identify a new cluster before an alert is issued. Similarly, an alert could be issued only when a new cluster has been identified for all features. As used herein, identifying new clusters using clustering analysis is intended to encompass all such options.

[0086] Our method issues a warning if any one of the four candidate features in the global DPMM can be used to identify a new cluster.

[0087] Figures 17-26 show the same results as in Figures 5-12, but in color. The correspondence between color and grayscale keys can be derived from the keys in Figures 5 and 17.

[0088] For the global DPMM results, clusters are presented in different colors, and the start of each cluster is indicated by a vertical line in the same color as the cluster label, where the first cluster is started upon the arrival of the first data point according to a threshold. The threshold set, in this case 1, specifies how many data points a given cluster must have before it is identified as a new cluster. Because this is an unsupervised method and does not require training on similar data, the DPMM is restarted from cluster 1 for each new tool. For the local DPMM results, workpiece changes are indicated by a black dashed vertical line where the DPMM is reset to 1. The cluster colors are kept in the same order as the global DPMM, except that the cluster start lines are not shown for clarity of the drawing.

[0089] The following analysis applies equally to the results in Figures 5-12 as it does to Figures 17-26. These results show how cluster identification can be used as a trigger to issue alerts, which are informative about the state of the component.

[0090] In general, cluster 1 is usually the largest (even after subtracting the effects of confounding influences), suggesting that the energy of AE harmonics has not changed enough under the current hyperparameters to warrant new clusters. Physically, this may mean that the early wear evolution of the tool does not significantly affect the AE harmonics. In most cases, the energy of harmonics increases sharply near the end of tool life due to tool wear. Due to the large forces acting on the tool, tiny dislocations and microcracks at grain boundaries within the tool material emit stress waves while also altering the tool's wear profile. As a result, the chip generation process permanently changes, resulting in increased energy in the AE signal. In the global DPMM, these structural and conditional differences manifest themselves as non-Gaussian clusters, prompting the algorithm to separate them into smaller Gaussian distributions, resulting in an increased number of clusters near the end of tool life.

[0091] In effect, global cluster initiation may now be used as an indication of tool degradation to the operator, and the appearance of a new cluster can be used to trigger an inspection rather than an inspection after a set sliding distance. Cluster initiation can therefore be seen as a novel way of setting a detection threshold. Observations from tool inspection can be used to determine whether a tool has been damaged beyond a predetermined tolerance threshold. In the latter scenario, the new cluster can be treated as an undamaged condition, thus allowing future similar observations to be classified into the same cluster.

[0092] However, if a cluster start occurs at the time of workpiece change and tool measurement, the operator should refer to the local cluster start. If the tool may have been damaged at the beginning of the pass due to impact with the workpiece, the local cluster can detect the corresponding AE fluctuation. In other words, resetting the cluster at workpiece change essentially ensures that the local DPMM will only trigger a new cluster when a significant change is seen in the data during machining of each workpiece. Therefore, it can be assumed that the change is not due to any factor other than tool wear (and related effects such as increased temperature and force), since all other operational effects are held constant.

[0093] Results from hyperparameter tuning indicate that for this dataset, a small value of α can be used since the clusters are well separated. This result can be beneficial, since in practice, many clusters are formed as cluster initiation and are a warning for tool inspection. If many clusters are initiated, the number of tool inspections also increases, thereby reducing the time savings achievable with DPMM.

[0094] Some tools show that after the start of a new cluster, data may be grouped into a previous cluster. For example, for tool B3 in FIGS. 11 and 23, after the start of the third cluster (represented by points of interest 11.1 and 23.1) at approximately 4 km, some data is still assigned to clusters 1 and 2 (the data identified by points of interest 11.2 and 23.2 contain mostly data points that fall into clusters 1 and 2). One explanation for this behavior is that the cutting edge is changing. As the tool moves along the workpiece, it is constantly shedding material from the cutting edge. Some particles ejected from the tool are larger than others, resulting in temporary holes in the cutting edge and momentary changes in chip formation characteristics. The method of the present invention can implement a short delay before issuing a warning, and analysis of the data points continues after the identification of the new cluster to determine what proportion of subsequent data points will be determined to belong to the new cluster. A threshold proportion of data points that belong to a newly identified cluster may be introduced before a warning is issued.

[0095] As a result, DPMM clusters the data into a new class. However, as the tool moves along the workpiece after this event, it is possible for the tool to move smoothly over the hole and return to an edge similar to the previous cluster. At this point, the AE data behaves much like the previous cluster, resulting in the behavior shown.

[0096] Another reason for the clustering mentioned above can be due to increased temperatures at the cutting edge. In some cases, the resulting chips can wrap themselves around the workpiece and collect at the cutting edge. In this case, the temperature at the cutting edge can rise sharply, burning the chips around the tool. As a result, chip formation characteristics can change as the temperature of the workpiece and tool affects material behavior (and, consequently, the resulting AEs). This is one reason among many, and temperature is another parameter of measurement on which cluster identification can be used to issue warnings.

[0097] The return of a particular data point to an earlier cluster can be useful for operators monitoring tools using DPMM as described above, and this information can also be useful for research and development purposes. This is especially true near the end of tool life, where tool failure clusters back to cluster 1 or 2, suggesting that AE behaves similarly to the beginning of tool life. The reduction in tool-to-workpiece contact area may be one reason for this observation. To avoid unnecessary testing due to confounding effects, the threshold can be increased so that a larger number of data points are required before a new cluster begins. This is therefore advantageous not only for increasing the time during tool use, thereby reducing the useful tool life remaining upon tool removal, but also for directing further research into component wear estimation.

[0098] However, the threshold value must be carefully considered, as a larger value means that cluster initiation is less sensitive to tool damage; clusters will not initiate until the threshold is reached and the tool may be severely damaged. An example of this can be seen in Figures 12 and 24, where global cluster 2 initiates, likely due to a confounding influence, i.e., a change in the workpiece (marked in the figures as relevant points 12.1 and 24.1, respectively). As can be seen, global cluster 2 begins suddenly for all four features, i.e., F1–F4. However, at this point in the local DPMM, a second cluster also immediately initiates, suggesting that the data was affected during the pass (marked in the figures as points of interest 12.2 and 24.2). This cluster initiation may be due to the interaction of the workpiece with the tool when entering the cutting state, where the sudden contact affects the chip-generating characteristics of the tool. The onset of cluster 4 can most likely be attributed to a confounding effect (marked in the figure as points of interest 12.3 and 24.3). Again, in global cluster 4, the significant and instantaneous change in acoustic emissions that causes the cluster onset suggests a confounding effect. Here, this can be used as a warning to the operator to inspect global clusters 3, 5, and 6 for damage.

[0099] This parallel clustering method is better for tool damage detection problems than two-class approaches, such as outlier analysis, for a number of reasons: When performing outlier analysis, it is assumed that the data is sampled from a Gaussian distribution so that Gaussian statistics apply.

[0100] Deviations from the Gaussian distribution are seen as outliers, indicating abnormal data. As mentioned above, DPMM can automatically separate large non-Gaussian clusters into smaller Gaussian clusters, eliminating the need to set the number of clusters a priori and thus enabling repeated detection of abnormal data. By assigning data to multiple clusters, it is possible to continuously learn about the tool's degradation stage while also being robust to confounding influences. For example, the second cluster onset for tools A2 and A3 and the fourth cluster on tools B1 and B4 occur due to large fluctuations in AE energy compared to the previous data. As evident from tool wear inspection, these occurrences are not caused by damage. By using a parallel DPMM method, it should be possible to avoid operator intervention in these cases; the same cannot be said for outlier analysis.

[0101] Without the ability to make continuous wear measurements, it is difficult to directly verify that the DPMM clusters in a meaningful way. One way to partially assess the success of the DPMM is to consider trials in which anomalies or irregularities were noted. In the next section, we consider clustering of heterogeneous tools to explore this further.

[0102] Clustering of heterogeneous tools

[0103] Predictions of tool wear states for tools that are different from the tools in the dataset may result in a lower positive angle since the training set is not representative of the test set. Since DPMM does not require a training phase, we consider here the feasibility of detecting changes in AE for tools that behave differently from others.

[0104] Figures 13 and 14, as well as Figures 25 and 26, show the global and local DPMM results for tools A5 and B5, respectively. For comparison, the same features and hyperparameters were used for these tools and all other tools in each grade. In both cases, it is clear that the AE features behave differently from the other features in the dataset. As can be seen, the global DPMM detects cluster 3 at approximately 1.6 km and 3 km for tools A5 and B5, respectively, due to unknown behavior that causes AE energy to be high at these specified frequencies. Due to this increase in energy midway through the tool life, the global DPMM does not detect new clusters near approximately 2.7 km (A5) and 5.2 km (B5), even though the signal energy has increased compared to the beginning of the tool life. However, at this point, the local clustering reflects this increased energy by creating a second cluster that can be used for intervention and avoid machining with a broken tool. Since DPMM does not utilize data from other tools for training, it is able to cluster different data in a way that can avoid machining with broken tools. Clearly, DPMM is robust to a wide range of tool behavior, suggesting its suitability.

[0105] Guidelines for parallel DPMM to avoid tool breakage

[0106] To accurately identify when a tool is about to break and avoid false positives, the following guidelines for parallel DPMM are suggested: Global clustering should be preferred over local clustering, unless the global clusters are initialized at the time of the workpiece change. This is because global clustering takes into account the entire dataset and, as a result, is more susceptible to significant changes in the dataset. Clusters during workpiece changes should be handled carefully, as a workpiece change may create new clusters. If a global cluster is started at the time of a workpiece change and a local cluster is in the started state during that workpiece, the start of that local cluster should be used as a warning. If a global cluster is started at the time of the workpiece change, but no local clusters were started during the workpiece, the start of the global cluster should be ignored as a warning.

[0107] Design modifications of the DPMM method as applied to tooling.

[0108] Many design variations can be made to the detailed embodiment described above. As noted above, the method can be used with any component of a machine that is susceptible to wear, and any parameter can be measured and analyzed. Furthermore, the method itself can be modified according to any of the steps detailed below.

[0109] Because DPMM does not rely on learning damage labels, the data set can be of any size. Because larger data sets contain more information, frequently averaging AE data can increase the frequency at which data arrives for clustering. In the detailed example above, the data points represent one second of data consisting of one million raw AE data points. However, in alternative embodiments, each second of data can be represented by many more data points (i.e., one data point represents 1 / 10 of a second, increasing the amount of information input into DPMM by ten times), thus creating a more useful trend or cloud of data for clustering. As more data becomes available, DPMM may be able to begin clustering earlier than currently possible because the cloud can be better defined with higher resolution, reducing the time required for tool testing. The specific representation of data points described above is not limiting; the input data points may represent any number of seconds of measurement data containing any number of raw parameter measurement data points, with the appropriate number of seconds and number of raw parameter measurement data points depending on the component and parameters to be analyzed. The selection of such an appropriate number of seconds or an appropriate number of data points may be made by one skilled in the art and may be modified using preliminary testing and / or live results.

[0110] The number of frequency bins can be increased, thereby achieving a higher frequency resolution and possibly deriving more features with increasing intensity near tool damage. Four parameter bins (here, frequencies) are proposed as input or candidate features in a detailed embodiment, but the number of input features may be 1-100, 2-50, 3-25, and preferably 4.

[0111] In this case, DPMM is used as an unsupervised technique where data is clustered without the need for a target value. In the future, DPMM can also be used as a semi-supervised learning technique where damage labels may be included in the model upon change detection.

[0112] Of interest in this study is that the increase in AE energy (harmonics) that leads to the final cluster initiation before failure typically occurs during the last four passes of the workpiece. Further investigation of this information may allow for the creation of a predictive system, and it may be possible to predict the remaining useful life of a tool by studying the nature of the final cluster initiation before severe damage.

[0113] Transfer learning can be applied to this dataset, where the transfer learning clusters (with known labels) are projected into a space that already contains unlabeled clusters. The idea is to minimize the distance between cluster means. For the dataset collected here, inductive transfer learning can be applied, where the algorithm has information about the current clusters and needs only a small amount of data from the new clusters to make similarity decisions. However, it is not clear whether this method is robust to tool breakage or tool wear.

[0114] The method of the present invention is compatible with dedicated microscopes that can capture 3D images while the tool is still attached to the holder, without the need to remove the tool. This eliminates positional changes that act as confounding influences in the current data set. Therefore, the relative importance and need for parallel DPMMs is reduced, and a single DPMM can be accurately employed.

[0115] Furthermore, for example, such a setup can be programmed with tool wear tolerances and thresholds that can be used to approach the tool at the start of the cluster, thereby eliminating the need for operator intervention and resulting in fully online tool wear testing.

[0116] conclusion

[0117] This study explores the concept of unsupervised learning, which can detect changes in monitoring data with an incomplete set of damage labels without requiring a training phase. DPMM and Gibbs sampling were used in this study to cluster AE data collected during the turning process. Because this method can be considered as an infinite Gaussian mixture model, the number of clusters does not need to be set a priori, resulting in a fully online implementation of the algorithm. By using unsupervised cluster onset as a warning / prompt for the operator to conduct a tool condition investigation, successful use of DPMM can reduce the time spent measuring the tool and provide a point at which the tool can be discarded. This method of using cluster onset as a detection threshold has not been seen in previous literature.

[0118] The input features for the model were taken from the frequency domain of the acoustic emission signals, which were measured over the life of the tool. Harmonics of chip-generating frequency energy were used as input features because they are generally found to increase in intensity as the tool approaches failure.

[0119] Throughout the detailed embodiments described herein, the clustering analysis used is DPMM, but it will be understood that other clustering analyses can be employed with the objective of achieving similar useful results. Other clustering algorithms that can identify clusters in the data related to the wear state of components include, but are not limited to, affinity propagation, k-means clustering, mean shift, agglomerative clustering, and spectral clustering.

[0120] Parallel DPMM saves time

[0121] Due to confounding influences affecting this data set, a combination of global and local DPMMs operating in parallel was implemented here for the first time in this paper. The global DPMM samples from the entire data set collected during the tool's life, while the local DPMM samples only from data collected during each workpiece (when the process structure changes). Using these parallel DPMMs, it is possible to reduce the time spent measuring tools while avoiding machining with broken tools. A complete list of time savings is provided in Table 1. Promising initial results indicate that tool measurement time can be significantly reduced from 42% of the average test time using this method to 13.2%. In this case, results are also compared to a preventive maintenance strategy, in which the operator stops machining when the tool reaches a predefined threshold (determined depending on the machining operation).

[0122] The results presented in this study are likely beneficial for industrial applications for a number of reasons. First, this method does not require training data that must include all operating conditions to safely predict unknown tools. Second, this method reduces the number of periodic tool wear measurements currently performed and allows damage labels to be collected only when prompted by the algorithm. By running DPMM in parallel, confounding influences can be avoided. This method also works with AE features that can be used online without requiring preprocessing, such as dimensionality reduction. Furthermore, because DPMM clusters according to feature variations, it is also possible to detect damage in tools containing various material compositions and even in tools that behave differently from other tools. In conclusion, this method can be applied to any tool, as long as the input features become more significant as the tool approaches failure. In fact, this method can be applied to any machine component and any input features (i.e., measurable parameters) that change as the component wears with use. Measurement time (min) Table 1: Summary of time savings using DPMM compared to other monitoring strategies JPEG0007749686000005.jpg84168Since the run-to-failure method accounts for over 42% of the test time for tool measurement, DPMM offers significant time savings.

[0123] Effect of hyperparameters

[0124] In the example provided herein, the hyperparameter α is fixed a priori at 20; a low value of α limits the number of clusters initiated. Tuning the hyperparameter α and threshold allows for the number of samples in each cluster to be varied. For general SHM, the presence of cracks is treated as damage, and cross-validation can be used with a false negative rate to check for misclassification of the DPMM, aiding in tuning the hyperparameter α. If the clusters are well separated, the number of resulting clusters is not affected by the α parameter.

[0125] When DPMM is used in this study to cluster the data online, it is not possible to learn the value of α a priori to find the optimal distribution. To understand the effect of the α parameter, a priori learning is required to find the optimal distribution. To understand the effect of the α parameter, the features used to obtain the results for tool A4 and tool B3 (Figures 8 and 11, respectively) were fed into the algorithm for various values ​​of α. In each case, i.e., for each α value, the algorithm was run 100 times to obtain the average number of clusters initiated. The results are presented using a boxplot in Figure 15. The average number of clusters over the 100 runs is shown as a red line, and outliers are indicated by red crosses that deviate beyond ±2.7σ (annotated by whiskers). The top and bottom of the box represent the 75th and 25th percentiles, respectively.

[0126] As can be seen from Figure 15, the number of clusters is independent of the value of α when α is small. This means that the clusters are well separated. When considering α values ​​between 0.1 and 50, the change in the average number of clusters is very small. Therefore, it is recommended in this case to use any value of α between 1 and 50, which does not significantly affect the number of clusters generated.

[0127] The window length of the Gibbs sampler can also affect the results of DPMM. Although the window length is not a hyperparameter, it is a value that can affect the number of clusters and the cluster initiation time. This is because the window length determines the number of data points that are reevaluated by the algorithm. In practice, this is not beneficial to the operator if the Gibbs sampler reevaluates data points from an early sliding distance and reassigns them to a different cluster than the one that exists at that time.

[0128] However, the window length needs to be large enough to cause the Markov chain to converge to the target distribution. Nevertheless, increasing the window length decreases the computation speed, so again, engineering judgment is required to find a reasonable window length.

[0129] Example System

[0130] 16 illustrates a system 100 according to an embodiment of the present invention. The system 100 includes an associated machine 102, a component 104, a sensor 106, a processing unit 108, and a mobile device 110.

[0131] Although the associated machine 102 is shown as having a component 104 and a sensor 106, this is by way of example only. The component 104 is removable and replaceable, which is advantageous for assessing wear. The sensor 106 may be provided within the associated machine 102, or may be additional and / or separate. In some embodiments, the component 104 includes the sensor 106. This may be advantageous when physical contact between the component 104 and the sensor 106 facilitates measurement of a relevant parameter during use of the component 104 (e.g., when forces exerted by or on the component 104 are measured). The sensor 106 may be a multi-purpose sensor configured to measure a variety of parameters during use of the component 104, or may be a sensor dedicated to a single parameter, such as acoustic emission. The sensor 106 may be configured to measure the parameter of the component 104 indirectly, i.e., without interfering with use of the component 104.

[0132] The processing unit 108 is any computer processor suitable for carrying out the steps of the inventive methods described herein. The processing unit 108 may be located within the associated machine 102, and may be provided with a wired connection, or may be provided with a wireless connection. In the illustrated embodiment, the processing unit 108 receives, during use, measurements 112 of parameters of the component 104 from the sensors 106 via a wireless connection. Suitable wireless communication protocols for such transmission will be apparent to those skilled in the art. Additionally, the measurements 112 may be sent to the processing unit 108 via the internet.

[0133] The processing device 108 performs at least one clustering analysis in accordance with an embodiment of the present invention. Measurements 112 may be received continuously from the sensor 106 or may be received in one or more data packets. The data processing device 108 performs one or more clustering analyses using the received measurements 112 and issues an alert 114 if it identifies a new cluster in the measurements 112. The alert 114 may be processed by the data processing device 108 itself and may include one or more of a computer-generated instruction to send to the associated machine 102 to terminate operation and a recommendation for component inspection or replacement for user confirmation. Alternatively or additionally, the data processing device 108 may send the alert 114 to a third device, such as a mobile device 110, for user confirmation. The data processing device 108 may be provided within the mobile device 110 or, as shown, may be separate therefrom. The communication between the processing unit 108 and the mobile device 110 may be wired or wireless, as described in connection with the communication between the sensor 106 and the processing unit 108 .

Claims

1. 1. A computer-implemented method for estimating wear of a component, comprising: performing a first clustering analysis on a first plurality of measurements to identify one or more clusters in the first plurality of measurements, the first plurality of measurements comprising measurements of parameters of the component taken while the component is in use; issuing an alert if a new cluster is identified in the first plurality of measurements; performing a second clustering analysis on a second plurality of measurements to identify one or more clusters in the second plurality of measurements, the second plurality of measurements including measurements of parameters of the component taken during further use of the component after the alert has been generated; issuing an alert if a new cluster is identified in the second plurality of measurements; continuing the first clustering analysis on the first plurality of measurements and introducing the second plurality of measurements into the first clustering analysis to identify one or more clusters in the combined first and second plurality of measurements; issuing an alert if a new cluster is identified in the combined first and second plurality of measurements.

2. The method of claim 1 , wherein the first and / or second clustering analysis is performed by an unsupervised machine learning algorithm.

3. The method of claim 2 , wherein the unsupervised machine learning algorithm does not prescribe the number of clusters determined during the first and / or second clustering analyses.

4. The method of claim 3 , wherein the unsupervised machine learning algorithm is a Dirichlet process mixture model.

5. The method of claim 1 , wherein the warning is a recommendation to perform a component inspection.

6. The method of claim 1 , wherein the parameter of the component is one or more of a temperature of the component, a force applied to the component, and an acoustic emission from the component.

7. The method of claim 6 , wherein the parameter is an acoustic emission from the component, and the first and / or second clustering analysis uses a plurality of adjacent frequency ranges as candidate features.

8. The method of claim 7 , wherein the component is a machining tool.

9. The method of claim 8 , wherein each frequency range corresponds to a predetermined harmonic of chip formation in the machining tool, and optionally each frequency range is spaced 4 kHz apart.

10. 1. A method for inspecting a component for wear, comprising: Implementing the computer-implemented method of claim 1; and in response to a warning output from said computer-implemented method, inspecting said component for wear.

11. 10. A data processing apparatus, comprising means for performing the steps of the method of claim 1.

12. A computer program product having instructions which, when executed by a computer, cause the computer to perform the steps of the method of claim 1.

13. A computer-readable storage medium having instructions embodied thereon that, when executed by a computer, cause the computer to perform the steps of the method of claim 1.

Images